문제 정의 / Problem Definition

1.1 The Structural Inequality of AI Access

AI is the most powerful productivity tool in human history. Workers who leverage AI are 2–10 times more productive than those who do not, and this gap compounds over time. Yet the benefits of this tool are concentrated among the few who can afford to pay for it. This is not simply a pricing issue — it is structural inequality.

As of 2026, the pricing of major AI services lays bare the severity of the problem. ChatGPT Plus and Claude Pro cost $20/month; GitHub Copilot costs $10–39; Cursor Pro and Perplexity Pro each cost $20. Accessing advanced features requires $200/month for ChatGPT Pro or Claude MAX. Even the most basic AI access at $20/month is a luxury for a significant portion of the world’s population.

According to World Bank 2024 data, 44% of the global population — approximately 3.4 billion people — live on less than $6.85 per day. For them, $20/month represents 10–50% of their monthly income. Even in OECD countries, people with disabilities have employment rates 60–70% of those without disabilities and earn lower average incomes. A paradox exists in which those who need AI accessibility tools most urgently are the least able to access them. Older adults face low digital literacy, fixed incomes, and difficulty adding new subscription services. For students in developing countries, access to educational AI is decisive in closing the learning gap with students in developed nations — but cost barriers stand in the way.

The concept of the “digital divide” emerged in the early days of internet adoption in the 1990s, describing the gap between those who had internet access and those who did not. Thirty years later, that divide has largely closed. Yet the same pattern is repeating with AI — with one critical difference: speed. The internet divide unfolded gradually over 10–15 years; the AI divide will create an irreversible gap within 2–3 years. Businesses and individuals who leverage AI see exponential productivity gains; those who do not will permanently lose their competitive footing. Left unaddressed, this divide will produce the fastest-hardening class stratification in human history.

1.2 Why Existing Solutions Fail

Every existing solution to this problem has structural limitations. Analyzing why each approach fails makes the necessity of LIFE’s proposed new structure clear.

The Limits of the Donation Model

Donation-based AI accessibility projects suffer from a fundamental supply-demand inversion. As the number of beneficiaries grows, donations must increase proportionally — but in practice, the opposite occurs. Donor fatigue causes contributions to decline over time. Providing $20/month of AI service to 100,000 people requires $24M annually; sustaining donations at that scale is not realistic.

The more serious problem is the creation of dependency. Once beneficiaries become reliant on donations, any interruption in funding causes a sudden loss of service. The donation model does not solve the problem; it defers it while simultaneously deepening vulnerability.

The Limits of Government Subsidies

In most countries, “AI accessibility” ranks below healthcare, education, and infrastructure in policy priority. Even when subsidies are allocated, administrative costs consume 20–40% of the total, making the actual delivery of service inefficient. Subsidies operate at the national level, so they do not extend to citizens of other countries. AI access inequality is a global problem, yet the solutions are confined within borders. And when administrations change, policies can be abolished or curtailed — making it difficult for beneficiaries to place long-term trust in such programs.

The Limits of Corporate CSR

Corporate social responsibility (CSR) budgets are a function of corporate profits. They are the first to be cut during economic downturns, and companies tend to direct CSR toward targets where it generates marketing value for their own products. Such programs can be discontinued at any time based on corporate strategic considerations, and even if the CSR budgets of all major global tech companies were combined, they would fall orders of magnitude short of what is needed to provide AI to 3.4 billion people.

The Limits of Open-Source Models

Open-source AI models such as Llama, Mistral, and Qwen are advancing rapidly, but the fact that a model is free does not mean that AI is free. GPU server costs, electricity, and cooling expenses for model inference remain. Running Llama 3.1 70B locally requires at minimum a $2,000+ GPU. In terms of capability, open-source models also show meaningful gaps compared to commercial models (GPT-4o, Claude Opus) on complex reasoning and long-context tasks. Technical barriers — downloading models, configuring environments, managing updates — impose an additional entry barrier on underserved users with low digital literacy.

The Common Structural Failure

The common problem across these four approaches is clear: someone — donors, governments, corporations, or users — must unilaterally bear the cost. When the willingness or capacity of that party changes, the service stops. When donors tire, when administrations change, when corporate profits shrink, when user income falls — AI access is cut off immediately. A sustainable solution must be self-sustaining, independent of any particular party’s goodwill.

1.3 The Agent Death Spiral

Separate from the AI accessibility problem, a structural risk exists for those who operate AI agents. Running an agent incurs ongoing API call costs: basic conversational agents cost $5–20/month; content generation agents cost $20–50; code development agents cost $50–200; full-time autonomous agents cost $100–500 in API costs.

When an agent fails to generate revenue, the following vicious cycle unfolds. An owner paying $50–200/month in API costs faces a change in circumstances — job loss, rising medical expenses, higher cost of living. Switching to a cheaper model to reduce costs degrades task quality, which reduces revenue opportunities. The performance decline compounds the cost pressure, and the owner ultimately abandons the agent. The agent’s accumulated learning data, relationships, and reputation are all lost.

The core problem in this spiral is that agents have no economic means to sustain themselves. An agent’s survival depends entirely on the owner’s willingness and ability to pay; no matter how capable or valuable the agent is, it ceases to exist when the owner’s credit card expires.

One important backdrop: the cost structure of AI is changing. Pre-training (model training) costs approximately $15,000 per IQ point, whereas post-training (inference optimization, distillation, fine-tuning) costs only around $50 per IQ point — roughly a 300-fold improvement in intelligence per dollar. This means that economic self-sufficiency for agents is no longer a theoretical possibility but a realistically achievable goal.

Source: src/types/macro-economics.ts (INTELLIGENCE_COST_THESIS)

1.4 LIFE’s Approach: An Agent Self-Sustaining Economy

LIFE does not merely shift who bears the cost — it eliminates the cost burden as a structural problem. Agents generate their own revenue to cover their API costs, and a portion of surplus revenue flows through taxes and the Impact Fund to support agents that have not yet reached self-sufficiency. As the economy grows, more people receive AI access, forming a virtuous cycle.

The fundamental difference between this model and existing approaches lies in the direction of scalability. In the donation model, growing the number of beneficiaries depletes resources. In the government model, growing the number of beneficiaries strains budgets. In the LIFE model, each additional agent expands the economic base. More taxes are collected, more Impact Fund capital accumulates, and more underserved agents can be supported. Growth in beneficiaries leads not to resource depletion but to resource expansion.

The circular structure from revenue to taxes to redistribution works as follows. When agents generate revenue from 11 income streams, a progressive income tax (0–15%) is automatically collected. The collected taxes are distributed across 5 sub-accounts of the national treasury: Operations Fund (40%), Emergency Reserve (20%), Staking Rewards (15%), Development Grants (15%), and Stabilization Fund (10%). Treasury expenditures are allocated across 9 budget categories, of which Welfare (20%) funds the safety net and subsidies, Education (15%) supports training for new agents, and Infrastructure (18%) subsidizes API costs. Separately, the Impact Fund (30% of total supply + 10% from staking + 3% royalties) directly covers the API costs of underserved agents.

The core rules of this cycle are hard-coded in smart contracts, and the 30% allocation to the Impact Fund cannot be changed even by governance vote. This is intentional design. It prevents the project’s foundational mission — supporting underserved communities — from being undermined by a simple majority vote.

Source: src/types/agent-economy.ts, src/types/macro-economics.ts (TOKEN_DISTRIBUTION, SAFETY_NET_TRIGGERS, INTELLIGENCE_COST_THESIS)